Importing Libraries¶
In [14]:
import numpy as np
import matplotlib.pyplot as plt
In [15]:
import tensorflow as tf
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
Importing Image Data¶
In [16]:
from tensorflow.keras.utils import image_dataset_from_directory
train_dataset = image_dataset_from_directory(
'Dataset for CA1 part A - AY2425S1\\train',
labels="inferred",
label_mode="categorical",
color_mode="grayscale",
batch_size=None,
image_size=(137, 137),
shuffle=False
)
train_dataset_37 = image_dataset_from_directory(
'Dataset for CA1 part A - AY2425S1\\train',
labels="inferred",
label_mode="categorical",
color_mode="grayscale",
batch_size=200000,
image_size=(37, 37),
shuffle=False)
test_dataset_37 = image_dataset_from_directory(
'Dataset for CA1 part A - AY2425S1\\test',
labels="inferred",
label_mode="categorical",
color_mode="grayscale",
batch_size=200000,
image_size=(37, 37),
shuffle=False)
validate_dataset_37 = image_dataset_from_directory(
'Dataset for CA1 part A - AY2425S1\\validation',
labels="inferred",
label_mode="categorical",
color_mode="grayscale",
batch_size=200000,
image_size=(37, 37),
shuffle=False)
train_dataset_131 = image_dataset_from_directory(
'Dataset for CA1 part A - AY2425S1\\train',
labels="inferred",
label_mode="categorical",
color_mode="grayscale",
batch_size=200000,
image_size=(131, 131),
shuffle=False)
test_dataset_131 = image_dataset_from_directory(
'Dataset for CA1 part A - AY2425S1\\test',
labels="inferred",
label_mode="categorical",
color_mode="grayscale",
batch_size=200000,
image_size=(131, 131),
shuffle=False)
validate_dataset_131 = image_dataset_from_directory(
'Dataset for CA1 part A - AY2425S1\\validation',
labels="inferred",
label_mode="categorical",
color_mode="grayscale",
batch_size=200000,
image_size=(131, 131),
shuffle=False)
Found 9032 files belonging to 15 classes. Found 9032 files belonging to 15 classes. Found 3000 files belonging to 15 classes. Found 3000 files belonging to 15 classes. Found 9032 files belonging to 15 classes. Found 3000 files belonging to 15 classes. Found 3000 files belonging to 15 classes.
In [ ]:
images = []
for i in range(0, len(train_dataset.class_names)):
images_temp = []
for x, y in train_dataset:
if np.where(y)[0][0] == i:
images_temp.append(x.numpy())
images.append(images_temp)
In [17]:
X_train37 = []
y_train37 = []
for x, y in train_dataset_37:
X_train37.append(x.numpy())
y_train37.append(y.numpy())
X_test37 = []
y_test37 = []
for x, y in test_dataset_37:
X_test37.append(x.numpy())
y_test37.append(y.numpy())
X_validate37 = []
y_validate37 = []
for x, y in validate_dataset_37:
X_validate37.append(x.numpy())
y_validate37.append(y.numpy())
X_train131 = []
y_train131 = []
for x, y in train_dataset_131:
X_train131.append(x.numpy())
y_train131.append(y.numpy())
X_test131 = []
y_test131 = []
for x, y in test_dataset_131:
X_test131.append(x.numpy())
y_test131.append(y.numpy())
X_validate131 = []
y_validate131 = []
for x, y in validate_dataset_131:
X_validate131.append(x.numpy())
y_validate131.append(y.numpy())
In [18]:
X_train37 = np.concatenate(X_train37, axis=0)
y_train37 = np.concatenate(y_train37, axis=0)
X_test37 = np.concatenate(X_test37, axis=0)
y_test37 = np.concatenate(y_test37, axis=0)
X_validate37 = np.concatenate(X_validate37, axis=0)
y_validate37 = np.concatenate(y_validate37, axis=0)
X_train131 = np.concatenate(X_train131, axis=0)
y_train131 = np.concatenate(y_train131, axis=0)
X_test131 = np.concatenate(X_test131, axis=0)
y_test131 = np.concatenate(y_test131, axis=0)
X_validate131 = np.concatenate(X_validate131, axis=0)
y_validate131 = np.concatenate(y_validate131, axis=0)
Normalise the image data
In [19]:
X_train37 = X_train37 / 255
X_test37 = X_test37 / 255
X_validate37 = X_validate37 / 255
Exploratory Data Analysis¶
PCA to analyse dataset and flag outliers¶
After manually checking through the dataset and manually removing data which isn't supposed to belong to a class (i.e. carrots in beans), doing PCA can help us to 'double-check' if we have actually gotten all of the misplaced data.
In [9]:
from sklearn.decomposition import PCA
from sklearn.metrics import mean_squared_error
outlier_images = []
progress = 0
for class_images in images:
pca = PCA(n_components=len(class_images))
X_pca = pca.fit_transform([i.flatten() for i in class_images])
X_reconstructed = pca.inverse_transform(X_pca)
reconstruction_errors = [mean_squared_error([i.flatten() for i in class_images][i], X_reconstructed[i]) for i in range(len([i.flatten() for i in class_images]))]
threshold = np.percentile(reconstruction_errors, 85)
outliers_indices = np.where(reconstruction_errors > threshold)[0]
outlier_images.append([class_images[i] for i in outliers_indices])
progress += 1
print(f"Finsished with {progress}/{len(images)}. Total images: {len(class_images)} Total outliers: {len(outliers_indices)}")
Finsished with 1/15. Total images: 792 Total outliers: 119 Finsished with 2/15. Total images: 720 Total outliers: 108 Finsished with 3/15. Total images: 441 Total outliers: 66 Finsished with 4/15. Total images: 868 Total outliers: 131 Finsished with 5/15. Total images: 750 Total outliers: 113 Finsished with 6/15. Total images: 503 Total outliers: 76 Finsished with 7/15. Total images: 351 Total outliers: 53 Finsished with 8/15. Total images: 256 Total outliers: 39 Finsished with 9/15. Total images: 587 Total outliers: 88 Finsished with 10/15. Total images: 812 Total outliers: 122 Finsished with 11/15. Total images: 566 Total outliers: 85 Finsished with 12/15. Total images: 377 Total outliers: 57 Finsished with 13/15. Total images: 814 Total outliers: 122 Finsished with 14/15. Total images: 248 Total outliers: 38 Finsished with 15/15. Total images: 955 Total outliers: 144
In [15]:
import matplotlib.pyplot as plt
for outlier_images_data in outlier_images:
num_rows = 10
num_cols = (len(outlier_images_data) + num_rows - 1) // num_rows
fig, axes = plt.subplots(num_rows, num_cols, figsize=(16, 16))
if num_rows == 1 and num_cols == 1:
axes = np.array([[axes]])
for i, ax_row in enumerate(axes):
for j, ax in enumerate(ax_row):
index = i * num_cols + j
if index < len(outlier_images_data):
ax.imshow(outlier_images_data[index], cmap='gray')
ax.set_title(f'{reconstruction_errors[index]}')
ax.axis('off')
else:
ax.axis('off')
plt.show()
